Feature
Research Report
The AI Transformation 100: Concrete Ideas for Fixing How We Work
There is one thing that is certain about AI: It is a hungry beast. It takes power to build and train large language models (LLMs). It takes processing speed to deliver generative AI that can answer complex questions, do in-depth research and — even — hallucinate wild answers to persistent questions at the speed of thought.
All types of AI eat massive amounts of high-speed computing power for lunch — and for breakfast and dinner. That means that at the core of everything, AI is a powerful processor, a chip created from silicon and innovation. The companies that build the fastest and smallest chips, perhaps those that can deliver high-power computing without drinking the planet’s resources, will win the heart — and investment dollars — of this industry. But it takes time and a huge investment to build this technology. And the race is already hot.
AI’s ravenous appetite for compute power is pushing the edge of Moore’s Law and forcing chip designers and chipmakers to innovate in ways they haven’t had to in years. Silicon has been creating wealth for decades. But this is a new gold rush.
Here are some of the top chip companies competing in the AI market.
Table of Contents
- 1. NVIDIA
- 2. AMD
- 3. Google
- 4. Amazon
- 5. Microsoft
- 6. Intel
- 7. Meta
- 8. Qualcomm
- 9. Cerebras
- 10. Apple
- 11. OpenAI
- 12. Extropic
- 13. Broadcom
- 14. Fractile
1. NVIDIA
NVIDIA is a dominant force when it comes to AI chips. This dominance has shot the company to the top of the stock market and made it one of the world’s most valuable companies. In fiscal year 2026, NVIDIA reported revenue of $215.9 billion — up 65% year-over-year — cementing its status as the most valuable company on earth by market cap.
The company’s AI chips dominate the market and are used by everyone from car makers to tech and AI companies. Plus, the pipeline of "AI factory" builds continues to thicken.
In 2025, NVIDIA announced the NVIDIA GB300 NVL72 rack-scale system, which is powered by the NVIDIA Blackwell Ultra architecture. GB300 systems posted record results in MLPerf Inference v5.1’s new reasoning benchmark, with the GB300 NVL72 showing about a 45% throughput gain over comparable GB200 racks on DeepSeek-R1.
If Blackwell promised a 25x reduction in inference cost and energy, Blackwell Ultra reads like the second act: 72 GPUs and 36 Grace CPUs in a fully liquid-cooled NVL72 rack, tuned for gigantic reasoning models and wired with Quantum-X800 or Spectrum-X fabrics to push output to industrial levels. The shape of the product makes the strategy obvious. NVIDIA is selling scaled systems for scaled models, and the market is responding in kind.
NVIDIA Acquires Groq's Assets
In late 2025, NVIDIA acquired Groq's assets for $20 billion, one of the largest acquisitions in AI hardware history. Groq, Inc. (not to be confused with Grok the chatbot) was founded by former Google engineers with a singular goal of making AI inference faster than any GPU could manage. The company’s Language Processing Unit (LPU) delivered deterministic, low-latency token generation that left traditional GPU inference looking sluggish by comparison.
The deal brought Groq's LPU architecture under NVIDIA's umbrella, and the company has since announced plans for Groq-3-based products. NVIDIA’s next-generation Rubin GPU architecture is also expected to launch in late 2026, promising another generational leap in AI compute performance.
2. AMD
NVIDIA is an AI-chip powerhouse, but long-time chip maker AMD is throwing some serious compute power into the game. Lisa Su, CEO of AMD, calls this a “ten-year AI cycle”, and the company is playing it long.
Last year, AMD announced the Instinct MI350 series, built on TSMC’s 3 nm process, with new low-precision formats like FP4 and FP6. This also coincided with a roadmap toward the MI400 Helios racks that aim to pull AMD further into hyperscale. AMD’s Helios is a massive, rack-scale AI computing platform designed to directly compete with NVIDIA's flagship NVL systems.
At CES 2026, AMD provided an early look at Helios and, for the first time, unveiled the full AMD Instinct MI400 Series accelerator product portfolio while previewing the next-generation MI500 Series GPUs. The latest addition to the MI400 Series is the AMD Instinct MI440X GPU, designed for on-premises enterprise AI deployments. The MI440X will power scalable training, fine-tuning and inference workloads in a compact, eight-GPU form factor that integrates seamlessly into existing infrastructure. It is slated to launch in the second half of 2026, anchoring AMD's push into massive "yotta-scale" cloud and data center environments
If you scan the world's fastest supercomputers, AMD silicon is everywhere. Frontier and El Capitan hold the #1 and #2 slots, proof of its ability to scale compute at absurd levels. Marketwise, AMD is still framed as number two, a role that carries pressure and freedom. Each new MI part plays as an invitation for hyperscalers who want a choice, and in that sense the finish line is beside the point. The company is building the stamina to keep running even as the course bends into unknown terrain.
Related Article: The Chips Cold War: How GPUs Became the World’s Most Valuable Political Resource
3. Google
By the end of 2024, Google’s sixth-generation TPU, Trillium, slipped into general availability like an actor finally stepping out from backstage. The chip was pitched as the workhorse behind Gemini training and then quickly absorbed into the mythology of Google Cloud’s “AI Hypercomputer.”
Trillium raised throughput by a factor of four compared with its predecessor, pulled efficiency up by two-thirds, and pushed memory to 32 GB per chip. The design can scale into pods of 256 units, a machine room’s worth of silicon aimed directly at training models like Gemini 2.0.
In 2026, Google pushed further with its seventh-generation TPU, dubbed Ironwood, which it describes as "the first Google TPU for the age of inference." Broadcom — Google's key custom silicon partner — is ramping deliveries of Ironwood TPUs through 2026, with even stronger next-generation demand expected in 2027.
The combination of bigger pipes, denser memory and Google’s own AI stack is the reason TPU climbed from a side note to a podium finish in the chip race.
4. Amazon
December of 2024 was Amazon’s turn at the podium. The company unveiled Trainium2 inside Trn2 instances and racks of UltraServers, chips pitched as twice the punch of the original Trainium with a cleaner energy profile. Then, in April of 2025, Anthropic made the announcement feel real: hundreds of thousands of Trainium2 chips would drive Project Rainier, a mega-cluster built to train Claude at a scale that sounds suspiciously like science fiction.
AWS kept layering on details through 2025. A Trainium server carries 16 chips, and four of those combine into an UltraServer that can reach roughly 83 petaflops. Trn2 instances are rentable in EC2, and Claude 3.5 Haiku already runs on the hardware.
Trainium3 UltraServers are now available, and make it easy for customers to train and deploy AI models faster at lower cost. Amazon EC2 Trn3 UltraServers powered by AWS's first 3nm AI chip enable organizations to run ambitious AI training and inference workloads.
5. Microsoft
We all think of Microsoft as a software company, but it also delivers some major hardware. The Cobalt 100 VMs moved from announcement to reality in late 2024, and by May 2026, they were running in 32 regions worldwide, powering workloads from Teams to Microsoft Defender.
How Azure Cobalt 100 VMs are powering real-world solutions, delivering performance and efficiency results https://t.co/P32eWB3Uqe pic.twitter.com/tBIlmoR2tO
— Eric Berg - MVP (@ericberg_de) September 24, 2025
The other half of the story is Maia. The original Maia 100 announcement spelled out the vision: a chip designed for Azure itself, feeding directly into OpenAI models and Microsoft’s own AI services.
In early 2026, Microsoft announced Maia 200, a next‑generation, in‑house AI accelerator that delivers faster, more reliable and more energy‑efficient artificial intelligence within the Azure cloud.
The new chip is an important step in Microsoft’s long‑term strategy, which is geared around building and optimizing its own AI infrastructure.
Webinar
Jun
16
From Legacy to Launch-Ready: How Gainbridge Made Its Website a Marketing-Led Growth Engine
Join in to learn how a D2C annuity brand gave marketing full website ownership — without slowing down or risking compliance.
Register
Webinar
Jun
25
The Hidden Cost of Fragmented Customer Communication
Discover why growing businesses are rethinking the systems, workflows and communication habits shaping customer experience.
Register
Webinar
Jun
30
How Modern Marketing Is Exposing the Limits of Legacy CMS
Why marketers are rethinking CMS workflows that slow publishing, personalization and campaign execution.
Register
Webinar
Jul
14
Content Leaders Collective: Proving Content’s Business Impact
Join us as top content leaders look beyond the buzzwords to share how they actually prove ROI and scale what works.
Register
6. Intel
Intel’s centerpiece is Gaudi 3, which is now shipping inside racks from Dell, HPE and Supermicro. The chip delivers training throughput that doubles what Gaudi 2 could manage and brings efficiency numbers meant to entice enterprises that find NVIDIA both scarce and expensive.
Gaudi 3 appeared in full-blown AI Factory systems, the kind of branded infrastructure bundles that look less like servers and more like industrial appliances. Intel presents it as open by design, running PyTorch, TensorFlow and Hugging Face stacks without proprietary gravity. The strategy flows from a simple proposition. Build hardware strong enough to train and infer at scale, keep pricing competitive and rely on OEMs to deliver it into the hands of enterprises that want an alternative to CUDA’s orbit.
Intel and AI startup SambaNova Systems are also actively collaborating on high-performance, cost-efficient AI inference solutions, following a major strategic partnership established in early 2026. Instead of a full acquisition, Intel invested $35 million into SambaNova, bringing its total stake to 8.2% and officially passing US antitrust clearance.
Intel continues to bet on open ecosystems as its key differentiator in a market increasingly gravitating toward proprietary silicon.
7. Meta
Meta has always lived or died by the feed, and today that feed runs on homegrown silicon. The company’s second-generation Meta Training and Inference Accelerator (MTIA v2i) first showed up at ISCA 2025 in the form of an academic paper, a rare peek into how Facebook, Instagram and WhatsApp keep recommendation models humming at planetary scale.
The chip is tuned for inference efficiency, with a 3.5x performance boost over the first generation, which makes sense given the billions of small decisions Meta’s platforms serve each second. MTIA v2i improves throughput and trims power draw, cutting costs for the endless act of ranking posts, reels and ads.
This new MTIA chip can deliver 3.5x the dense compute performance & 7x the sparse compute performance of MTIA v1.
— AI at Meta (@AIatMeta) April 10, 2024
Its architecture is fundamentally focused on providing the right balance of compute, memory bandwidth & memory capacity for serving ranking & recommendation models. pic.twitter.com/8THnmi3JFD
Meta is also one of Broadcom's key XPU customers. In an April 2026 announcement, Meta began partnering with Broadcom to co-develop multiple generations of custom silicon that’ll ensure they (Meta) have the compute foundation to deliver on their long-term AI ambitions.
Broadcom now expects to deliver its first gigawatt of Meta compute in 2027, representing a $12–15 billion revenue opportunity.
8. Qualcomm
Qualcomm has built an empire on the idea that intelligence belongs in your pocket and in your hands just as much as in the cloud. In 2025 the company extended that logic into the laptop bag. The Snapdragon X Elite and X Plus began shipping inside Microsoft’s Copilot+ PCs, pulling Dell, Lenovo and Samsung along for the ride. What used to be a processor spec now feels like a ticket into a new category of computers, machines that wake up already speaking the language of multimodal AI.
Phones carried the same momentum. In September, Qualcomm announced the Snapdragon 8 Elite Gen 5, the chip that will sit beneath the glass of Android flagships. Its NPU is engineered to juggle chat assistants, image generation, translation and whatever other party tricks developers decide to cram into a handset.
In 2026, Qualcomm launched two all-new Snapdragon mobile platforms designed to deliver faster, smoother and more immersive mobile experiences to users worldwide. The Snapdragon 6 Gen 5 brings premium features to a wider selection of devices with new-in-series AI camera, gaming and performance functions as well as increased power efficiency.
The Snapdragon 4 Gen 5 also was unveiled which raises the bar for performance in its segment providing reliable connectivity, long-lasting battery life as well as gaming abilities.
Related Article: The Rise of AI Factories: Inside the New Data-to-Agent Pipelines
9. Cerebras
Cerebras has always chased size as a virtue. Its Wafer-Scale Engine 3 is the largest chip ever produced, carrying four trillion transistors across 900,000 cores. Instead of chopping a wafer into hundreds of dies, Cerebras keeps it whole, creating a slab of silicon that can hold entire models without splitting them across chips.
That approach now powers Condor Galaxy, a distributed AI supercomputer built with G42. The third system, CG-3, pushed the network to 16 exaFLOPs of capacity.
Cerebras made a landmark move in May 2026, completing the largest US tech IPO since Snowflake in 2020. The company raised $5.55 billion after pricing shares at $185 — well above its initial target range — with the stock nearly doubling on its NASDAQ debut under the ticker CBRS, giving it a market cap of roughly $95 billion at close.
Before debuting on NASDAQ, Cerebras had already secured a cloud deal with OpenAI worth more than $20 billion, and an AWS partnership to bring its chips into Amazon data centers. The company has also shifted its business model increasingly toward cloud services — going up against Google, Microsoft and Oracle — rather than pure hardware sales.
10. Apple
Apple might be late to the game compared to other chip companies here. But lagging behind everyone else has never stopped Apple from innovating new technologies — or markets.
In October, 2025, Apple officially announced its new silicon, the Apple M5. Built on a 3-nm process, it boasts a significant jump in AI performance, up to around 4x peak GPU compute for AI tasks compared to M4. The M5 features a 10-core GPU — with Neural Accelerators embedded in each core — and enhanced unified memory bandwidth (around 153 GB/s) to better handle on-device AI workloads.
The M5 chip is already being used in new devices, such as the 14-inch MacBook Pro, the updated iPad Pro and the upgraded Vision Pro headset. According to Apple, this chip marks the “next big leap in AI for the Mac.”
M5 is now part of the Apple Vision Pro
More recently, Apple has reportedly reached a preliminary agreement with Intel to manufacture some of its custom-designed chips. The deal is expected to diversify Apple’s manufacturing beyond longtime partner TSMC, especially as demand for AI chips continues to strain global semiconductor capacity.
Reports suggest Intel could initially produce lower-end or legacy iPhone, iPad and Mac processors using its advanced 18A process technology, while TSMC would continue handling Apple’s flagship silicon.
Beyond devices, Apple is also building for on-device AI and cloud/edge. The tech company is reportedly developing specialized chips for servers and smart glasses (in addition to Macs) to support its Apple Intelligence platform. The company also indicated it is exploring generative AI / chip-design automation, using AI to help design the silicon itself.
11. OpenAI
The ChatGPT-maker doesn't have anything on the market yet, but it's working on developing its own AI chips and systems in partnership with Broadcom. These chips are not for commercial sale — they will be used within OpenAI's own operations to improve efficiency and cut down on the compute demands of its AI models.
"By building our own chip, we can embed what we’ve learned from creating frontier models and products directly into the hardware, unlocking new levels of capability and intelligence," said OpenAI's co-founder and President, Greg Brockman.
OpenAI's first-generation custom chip — code-named Titan 1, co-developed with Broadcom and manufactured at TSMC — is now confirmed for volume deployment in 2027 at over one gigawatt of compute capacity.
The timeline has moved from the originally reported "late 2026" to 2027 as the scope of the program expanded. Analysts have projected that Titan 1 alone could contribute $10 billion or more in revenue to Broadcom by H1 2027, making it one of the largest custom silicon engagements in industry history.
12. Extropic
Extropic is taking a much different approach to AI hardware with what it calls the world’s first scalable probabilistic computer, a new class of AI chip that generates probability samples directly, instead of performing energy-intensive matrix math on GPUs.
Its chip, known as the Thermodynamic Sampling Unit (TSU), is built from arrays of transistor-based probabilistic bits, or pbits, which fluctuate between 1 and 0 and can be tuned to represent probability. By combining massive numbers of pbits, TSUs can sample from complex probability distributions with far less energy than digital processors. The goal, according to Extropic, is overcome the top barrier to major AI adoption: not enough electricity.
Extropic also introduced the Denoising Thermodynamic Model (DTM), a generative AI algorithm developed specifically for TSUs. According to the startup, simulations show that running DTMs on TSUs could be up to 10,000× more energy-efficient than current algorithms running on GPUs.
Extropic has produced a hardware proof of technology and a development platform called XTR-0. It has been beta tested by early partners, but is not yet a commercial-scale system.
13. Broadcom
Broadcom may not design chips for the open market the way NVIDIA does, but it has quietly become one of the most consequential forces in AI silicon — and the numbers make it impossible to ignore. Broadcom's model is built around custom silicon — specifically XPUs (accelerator processing units) and ASICs designed in close collaboration with hyperscalers. Its named XPU customers now include Google, Meta, ByteDance, Anthropic, Fujitsu and OpenAI, with each relationship representing three to five years of recurring silicon revenue.
CEO Hock Tan stated the company has a "clear line of sight" to over $100 billion in annual AI chip revenue by late 2027 or early 2028, backed by a $73 billion backlog and secured production capacity at TSMC for 3nm and 2nm nodes through the end of the decade.
Beyond custom chips, Broadcom also supplies the high-speed networking switches, including its Tomahawk 6 Ethernet switch, that tie AI accelerator clusters together. This combination of custom silicon and networking infrastructure has made Broadcom an indispensable partner for hyperscalers building Gigaclusters — data centers consuming over one gigawatt of power.
14. Fractile
Not every AI chip company is chasing the training market. Fractile, a UK-based startup founded in 2022 by Oxford-trained chip engineer Walter Goodwin, has its eyes fixed on a one problem: inference latency. The company’s recent $220M raise sheds spotlight on AI inference hardware as startups race to cut the cost, latency and power demands of frontier models.
Fractile's central insight is that the most advanced frontier models now require tens of millions of tokens to solve hard problems, and generating each token demands moving enormous amounts of data between processors and memory. That data movement creates latency, and latency is the enemy of real-world AI deployment at scale.
To address this, Fractile has developed a chip architecture that integrates memory directly within a standard server rack, dramatically reducing the distance data has to travel. Importantly, the design does not rely on traditional high-bandwidth memory (HBM) or on-chip SRAM — suggesting an entirely novel memory integration approach.
"Compressing a month of work into a day, a weekend of lab computation into a coffee break, will make all that work happen radically faster, but it will also make far more ambitious AI use cases economically viable," said Goodwin.